At the end of 2012 I decided to challenge theorists who talked about
“prediction” in AGI to get them to try using “prediction” in their real
life.  If prediction is just another name for “understanding” without
providing us with some greater insight into how understanding is attained
then what good is it?  So I decided to try using prediction in my real life
to see how it worked.  I made a prediction, based on my feelings, that I
would be able to create an AGi program within a year to a year and a
half.  AGi,
with a little ‘i’ is intended to represent some clear advancement in AGI
but something that is also clearly less than fully human level AGI.  I then
decided to use my prediction in a way that would be clear to everyone to
show that if prediction might lead to some understanding it would have to
do so through an application of a mature integrity of method.  So if my
predictions do not turn out to be reasonably good then I will need to
acknowledge that there was something wrong in theories that founded the
basis of my prediction.  If my predicted successes do not occur, there must
have been something wrong in the theories that led me to have the
confidence to make the prediction in the first place.  If you fail, then
you have to recognize the failure if you want to benefit from the
experience.  (That makes sense.)  Now some people made the claim that I did
not understand what they meant by their use of the word prediction in their
theories or that my attack was elementary.  And someone explained that the
kind of prediction that they were thinking about in general intelligence
occurs at the neural level and so (I figured they were saying) it is not a
conscious process that one can test on their own ego-driven assertions of
their ability to make an advancement in the field.  Well, I’m sorry but
those criticisms are just not good enough because they avoid the challenge
without any evidence that they had actually made an advancement that
transcended my skepticism.  My challenge was simple and something that was
feasible.  Those kinds of criticisms all look a little too familiar.  If
the opponent is much more naïve about the problem then you are then his
challenge can be dismissed out of hand.  There is no need to even think
about the criticism, much less take a challenge based on it. But it seems
obvious to me that you need some way to test your theories and if you can
never meet any of your projections of future accomplishment that were based
on those theories then they must be off the fringe.



One of the things I found in taking my own challenge was that it started to
lead to some realizations that predictions could be used as
benchmarks.  Furthermore
I discovered that the ‘predictions’ had a tendency to cause certain points
of view to crystallize in a way that I found that I had to deal with those
ideas more succinctly than I would have if those viewpoints had not been
crystalized.



Anyway, one of my original predictions was that it might take me 5 months
before I started testing my theories out, but, then I realized that if it
did take me that long then it would be a sign that something was seriously
wrong.  And indeed, during the past month, the 3rd month of my project, I
did not actually work on my program at all and I do not foresee having the
time to work on during the next month.  This is a serious failing on my
part, and even though I originally foresaw something like this happening,
my subsequent realization that the 3rd month was critical to meeting a time
table I can now conclude that there is absolutely no way I will meet my
original 2 year deadline.  It is an extremely negative indicator.



However there is one good indicator.  During the last few days I developed
a new theory that learned specialization is a necessary basis for learned
(or true) generalization.  While this may seem obvious, or perhaps a bit
glib, I believe that it may prove to be a major part of my theory.  It is a
little different than other theories by which higher general intelligence
might be based on specialization.  Minsky’s common sense reasoning, for
example, was founded on the principle that higher general intelligence
would be based on the knowledge of everyday facts.  One of Ben Goertzel’s
theories was that general AI could be founded on thousands of narrow AI
algorithms.  (I may not have expressed those theories in just the right way
but I think I was close.)  My theory is different because it is based on
the ability of an AGI program to learn.  If an AGI program can learn to
effectively utilize many specializations in its interactions with the IO
world, then it would be able to generalize that knowledge effectively just
because it would (by hypothesis) be able to use those generalities
effectively for many of the particular variations that might be needed.  The
ability to specialize is the both the basis for developing generalizations
and the basis for applying generalizations.



My use of prediction in my everyday life helped me to crystalize some
benchmarks to measure the progress (or the lack of it) I was making with my
project but it also helped me crystalize methods of analyzing why my
project did not progress as well as I hoped it might.  Similarly, I believe
that my theory that learned specialization is necessary for learned
generalization may help me crystalize stepped tests for AGi.  This is
important because so far it has seemed like the gradual development of an
AGI program has been methodically unsound.  If I can make careful tests,
starting out with something that I know is feasible and then gradually test
incremental improvements then there is a much greater chance that I will be
successful than if I had to wait until I had it all figured out.



So although my benchmarks of progress are extremely negative so far, I feel
as if I have a new idea that will help me get started once I find the time.



Jim Bromer



-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com

Reply via email to